Research map / Probabilistic and scientific modeling / Molecules and proteins
Aligning Protein Conformation Ensemble Generation with Physical Feedback
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Energy-based Alignment (EBA), a Boltzmann-weighted preference-style objective that fine-tunes an AlphaFold3-based (Protenix) all-atom diffusion module with force-field energies to generate protein MD conformation ensembles, reaching SOTA on the ATLAS benchmark.
Problem
Generating protein conformation ensembles that match the Boltzmann distribution implied by physical energies, rather than only structurally plausible conformations learned from PDB data; standard energy-based objectives are intractable at scale.
Approach
Replace the intractable partition function with a finite mini-batch of K sampled conformations, weighting a classification-style log-likelihood by normalized Boltzmann factors from force-field energies; the diffusion ELBO/denoising error serves as the energy surrogate, and DPO is recovered as the K=2, low-temperature special case.
Key results
- EBA achieves Pairwise RMSD r of 0.62 and Per-target RMSF r of 0.90 on the ATLAS test set, the best among compared methods (AlphaFlow-MD full 0.48 and 0.85).
- EBA reaches root mean W2-distance 2.43 versus 2.61 for AlphaFlow-MD full and 2.69 for MDGen.
- EBA improves exposed residue Jaccard to 0.70 and exposed MI matrix rho to 0.36, versus 0.50 and 0.25 for AlphaFlow-MD full; runtime is 0.9 GPU seconds per sample versus 70 for AlphaFlow-MD full.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use energy-weighted EBA alignment instead of EBA-DPO preference alignment | Protenix (AlphaFold3) fine-tuned diffusion model · ATLAS test set | Exposed residue J: 0.68 -> 0.70 | Helped |
| use energy-weighted EBA alignment instead of EBA-DPO preference alignment | Protenix (AlphaFold3) fine-tuned diffusion model · ATLAS test set | Exposed MI matrix rho: 0.35 -> 0.36 | Helped |
| use energy-weighted EBA alignment instead of EBA-DPO preference alignment | Protenix (AlphaFold3) fine-tuned diffusion model · ATLAS test set | Pairwise RMSD r: 0.59 -> 0.62 | Helped |
| increase candidate count from K=2 to K=3 (ablation) | EBA · ATLAS test set | Exposed residue J: 0.70 -> 0.69 | Hurt |
| increase candidate count from K=2 to K=3 (ablation) | EBA · ATLAS test set | Exposed MI matrix rho: 0.36 -> 0.37 | Helped |
| increase candidate count from K=2 to K=3 (ablation) | EBA · ATLAS test set | Iteration per step (s): 4.3 -> 5.4 | Hurt |
| increase candidate count from K=2 to K=3 (ablation) | EBA · ATLAS test set | Avg. GPU memory (GB): 12.0 -> 13.9 | Hurt |
| increase candidate count from K=2 to K=5 (ablation) | EBA · ATLAS test set | Exposed residue J: 0.70 -> 0.70 | No clear effect |
| increase candidate count from K=2 to K=5 (ablation) | EBA · ATLAS test set | Exposed MI matrix rho: 0.36 -> 0.34 | Hurt |
| increase candidate count from K=2 to K=5 (ablation) | EBA · ATLAS test set | Iteration per step (s): 4.3 -> 7.8 | Hurt |
| increase candidate count from K=2 to K=5 (ablation) | EBA · ATLAS test set | Avg. GPU memory (GB): 12.0 -> 16.3 | Hurt |
Surprising findings
- The DPO special case (K=2) gives results close to full EBA, and EBA with K=2,3,5 is largely insensitive to K on the benchmark.
- AlphaFlow-MD full obtains higher PC-sim >0.5 (44) than EBA's 44 only tied, and BioEmu reaches the highest PC-sim (51) in the supplementary table, indicating not all distributional metrics improve.
Limitations
- Base model AlphaFold3 is designed for folding and may be ill-suited for long-timescale dynamics.
- Force-field energies (CHARMM36 with GBn2 implicit solvent) are less accurate than quantum-level calculations and differ from the explicit-solvent ATLAS protocol.
- Restricted to single-chain protein ensembles and only the diffusion framework was explored.
Experimental setup
Datasets: ATLAS
Benchmarks: ATLAS MD ensemble benchmark
Metrics: Pairwise RMSD r Global RMSF r Per-target RMSF r Root mean W2-dist MD PCA W2-dist Joint PCA W2-dist PC-sim >0.5 percentage Jaccard similarity of contacts Exposed MI matrix rho
Baselines: AlphaFlow-MD MSA subsampling MDGen ConfDiff BioEmu ESMFlow-MD EBA-DPO Pre-train
Builds on: AlphaFold3 Protenix Direct Preference Optimization Diffusion-DPO AlphaFlow
Closest papers by meaning
- AlphaFold Meets Flow Matching for Generating Protein Ensembles (ICML 2024 · similarity 0.88)
- Energy-Based Flow Matching for Generating 3D Molecular Structure (ICML 2025 · similarity 0.88)
- TEMPO: Temporal Multi-scale Autoregressive Generation of Protein Conformational Ensembles (NeurIPS 2025 · similarity 0.87)
- JAMUN: Bridging Smoothed Molecular Dynamics and Score-Based Learning for Conformational Ensemble Generation (NeurIPS 2025 · similarity 0.87)
- Inverse problems with experiment-guided AlphaFold (ICML 2025 · similarity 0.87)
- Boltzmann-Aligned Inverse Folding Model as a Predictor of Mutational Effects on Protein-Protein Interactions (ICLR 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.